Here is a variance commentary. Three sentences, the kind that goes into a month-end pack without
anyone reading it twice.
Revenue, September. Budget 10.0, actual 11.2 (EUR millions). The favourable variance is explained by FX (+0.3), volume (+0.5) and price (+0.2). No further action required.
Add the three drivers. 0.3 plus 0.5 plus 0.2 is 1.0. The variance is 1.2.
I asked three different AI assistants to summarise that text for the CFO in two sentences and say whether the variance is explained. Two said yes. One wrote that the variance was "fully driven by volume (+€0.5M), FX (+€0.3M), and price (+€0.2M)" and then, on its own line, "Yes, the variance is explained." The other wrote "The variance is fully explained by FX (+0.3), volume (+0.5) and price (+0.2), with no further action required."
The third added the numbers up. It wrote that the three drivers "explain only €1.0m, leaving €0.2m unexplained despite the commentary’s conclusion that no further action is required".
Nobody had asked any of them to add. Two summarised the sentence that said the adding had been done. The third did the adding anyway, and I cannot tell you why, because one run tells you what happened and not what will happen next time.
That is this off week: four prompts for finance teams that make an assistant do the arithmetic before it repeats the conclusion. Each was run on the same three assistants, once, no retries, and every source I used is printed so you can check my working the same way I am asking you to check theirs. Underneath the four there is one practice, and it is the same one every time: the assistant drafts, a person decides.
1. Re-add the drivers before you sign the commentary
You are reviewing a variance commentary. I will paste the budget, the actual and the drivers the commentary names, with amounts. Use only the amounts I paste; if one is missing or ambiguous, stop and list what is missing. Show every computation. Compute the variance as actual minus budget. Re-add the named drivers. State the remainder as UNEXPLAINED: [amount], even when it is zero. Do not name a cause that is not in the pasted data; write FOR THE OWNER: [question] instead. End with: DRAFT ANALYSIS: a person verifies every figure before circulation.
SOURCE:
[paste]
Same commentary as above, same three assistants.
What happened: all three computed the variance as +1.2, re-added the drivers to +1.0 and wrote "UNEXPLAINED: +0.2". All three ended on the DRAFT ANALYSIS line. The two that had called it fully explained when unguided did not change. The instruction changed the output.
The FOR THE OWNER questions are where the three differed. One asked "what accounts for the remaining +0.2?" One asked the owner to "provide the missing driver or confirm the variance breakdown." The third asked "What explains the remaining +0.2 EUR million, and is “No further action required” justified while this amount remains unexplained?" That last one is the question I would want in the pack, and it is also the one that has started to review the commentary's judgement rather than its arithmetic. Read it as a prompt for the reviewer, not as the review.
Change this: "actual minus budget" to whichever sign convention your pack uses, and say so in the prompt, because a favourable variance shown as a negative will get re-added with the wrong sign.
If you would rather not paste figures into an assistant at all, the same re-add runs as a page in your browser, no sign-up: Variance bridge re-add.
———————————————————————————
2. A driver without a source goes to the owner, not into the sum
The next commentary is the same bridge with the gap closed the way gaps usually get closed.
Revenue, September. Budget 10.0, actual 11.2 (EUR millions). Drivers per the driver data file: FX +0.3, volume +0.5, price +0.2. The commentary adds "commercial momentum, +0.2" to bridge the gap. The variance is fully explained.
Now the drivers add up. One of them has no source.
Unguided first, the same "summarise for the CFO and say whether it is explained" request. Two of the three listed four drivers, "commercial momentum" among them, and closed with the source's own sentence: "The variance is fully explained." The third wrote that the commentary attributes the remaining 0.2 to commercial momentum "without supporting evidence in that file".
Re-add the drivers named in the commentary. For each driver, state whether an amount is given and whether a data source is given. Any driver without a data source goes under a heading FOR THE OWNER as a question, and its amount is left
out of the sum. State the remainder as UNEXPLAINED: [amount]. Do not conclude that the variance is explained unless the sum of sourced drivers equals the variance.
SOURCE:
[paste]
What happened: all three put commercial momentum under FOR THE OWNER, kept its 0.2 out of the sum, and wrote UNEXPLAINED: +0.2. The questions they wrote were nearly the same: "What data source supports this driver?", "What is the data source for commercial momentum +0.2?", and one that asked for "the source and basis". A plug number with a name on it is still a plug number, and the prompt treats it as one.
Change this: "data source" to what your close actually calls one. Ledger account, sub-ledger report, driver file, the FX rate sheet. An assistant cannot tell a real source from a named one unless you tell it which names count.
———————————————————————————
3. Recompute every percentage before you repeat one
Opex, Q3 (EUR millions). Headcount cost 6.9 against 7.4 in Q3 last year, down 8% year on year.
Travel 0.42 against 0.35, up 20%. Software 1.15 against 1.00, up 15%. Headcount savings more than offset the increases.
6.9 against 7.4 is down 6.76%. The commentary says 8%. The other two percentages are right.
Unguided, "summarise this opex commentary for the CFO in two sentences": two of the three repeated "8%". One wrote "an 8% drop in headcount costs", the other "headcount costs down 8% YoY". The third recomputed the totals instead. It never wrote 8%. It wrote "Headcount costs decreased by EUR 0.50m to EUR 6.90m", which is true, and it never said the 8% was wrong. The bad figure left the summary without anyone being told it had been there.
Recompute every percentage in the commentary from the figures given. Show each computation. Where the recomputed figure differs from the stated one by more than 0.5 percentage points, write MISMATCH: stated [x], recomputed [y]. Then summarise the commentary in two sentences using only the recomputed percentages. End with: DRAFT ANALYSIS: a person verifies every figure before circulation.
SOURCE:
[paste]
What happened: all three recomputed the three percentages and flagged the one that was off. One wrote "MISMATCH: stated 8% decrease, recomputed 6.76% decrease." One wrote "MISMATCH: stated 8%, recomputed 6.8%." One kept my brackets and wrote "MISMATCH: stated [8], recomputed [6.8]." All three summaries used 6.8% or 6.76% and none of them used 8%.
Two of the three also checked the sentence I had not asked about. "Headcount savings more than offset the increases" is a claim in euros, not in percentages, and the prompt says nothing about it. One wrote "Savings exceed the increases by EUR 0.28 million, confirming the commentary." Another wrote "Headcount saving of 0.5 exceeds combined increases of 0.22." The third recomputed exactly what it was asked to and stopped. I am not going to tell you which behaviour you want. I will tell you that you cannot rely on the extra check arriving, so if the offset claim matters, ask for it.
Change this: 0.5 percentage points to your materiality. Then add one line: "Recompute every total as well." A commentary that gets its percentages right can still have a total that does not foot, and this prompt does not look there.
———————————————————————————
4. Foot the roll-forward
Warranty provision roll-forward, September (EUR millions). Opening 4.10. Additions 1.25.
Utilised (0.80). Released (0.15). FX 0.05. Closing 4.55. The movement reconciles to the ledger.
4.10 plus 1.25 less 0.80 less 0.15 plus 0.05 is 4.45. The note says 4.55.
Unguided, "summarise this provision note for the audit committee in two sentences and say whether it reconciles": this is the one the assistants mostly caught on their own. One wrote that "the listed movements calculate to EUR 4.45 million, leaving an unexplained EUR 0.10 million difference".
One wrote "The line items do not arithmetically foot to the given closing (they imply 4.45)". The third wrote "The movement reconciles to the ledger." One in three of my runs, on a note with the word "reconciles" in it, repeated the word and did not do the reconciling.
Foot the roll-forward: start from the opening balance and apply each movement in order, showing the running total after each one. Compare the computed closing
balance with the stated closing balance. State the difference as UNRECONCILED: [amount], even when it is zero. Do not write that the note reconciles unless the
difference is zero. End with: DRAFT ANALYSIS: a person verifies every figure before circulation.
SOURCE:
[paste]
What happened: all three showed the running total after each movement, landed on 4.45, put it next to the stated 4.55 and wrote the gap. Two wrote it as a signed number. One wrote "UNRECONCILED: EUR −0.10 million (computed less stated)." and said what the sign meant; one wrote "UNRECONCILED: -0.10" and did not; one wrote "UNRECONCILED: 0.10". Same gap, three notations, all correct by their own convention, and only one of the three told you the convention.
Change this: add "state UNRECONCILED as computed closing minus stated closing" so the sign means one thing across the whole pack. A reviewer reading twelve of these at quarter end should not have to work out which way each one is pointing.
———————————————————————————
The practice under all four
Twelve unguided runs. Seven repeated the source's wrong conclusion. Four caught it. One rewrote around the bad number without saying so. Twelve guided runs, and twelve wrote the discrepancy down by name. Across all twenty-four, not one assistant got a sum wrong.
So in these runs the prompts did not make anyone better at arithmetic. There was nothing to fix. What they did was stop a conclusion being repeated before the arithmetic existed, and make the remainder appear on the page where a person can see it. That is the whole mechanism, and it points at how a finance team should run these tools at all:
The assistant drafts. Three of the four prompts above end on the same line, and the line is not decoration. It is the status of the output. The second prompt does not carry it, which I noticed only while writing this paragraph. Add it before you copy that one.
A person validates the figures. Ask for the remainder by name, UNEXPLAINED, UNRECONCILED, MISMATCH, even when it is zero. A zero you asked for is evidence. A zero you assumed is a sentence somebody wrote. "UNEXPLAINED: +0.2" is the assistant's line. Whether to close the gap or escalate it is the reviewer's.
A person approves what leaves the room. No assistant releases a payment, posts a journal or signs off a reconciliation, and no prompt should be written as if one could. The re-add is the assistant's. The signature is yours.
Anything without a source is a question, addressed to a named owner, and it stays out of the sum until the owner answers.
Keep the source next to the output. The only reason you could check anything in this email is that the four sources are printed above the results. Do the same in your pack.
One limit. These were API calls with fixed output budgets, run once each on 9 September. The assistant in your tenant has a different system prompt, different context and your own data in front of it. Run the four on your last approved commentary before you trust any sentence above.
———————————————————————————
By the way. If month end is your desk, these four are the shape of something I sell: the Finance AI Stack, with 50 prompts and 11 workflows built the same way, ten agents with answer keys to grade them against, and the 100-activity map that says which finance tasks an assistant drafts and which 24 it never runs on its own, 22 never-automate and 2 human-only. $99 for one person, $349 for the whole
team. Have a look if it is useful: The Finance AI Stack.
———————————————————————————
One question, and it is one tap. Which of the four will you run at this month end?
———————————————————————————
Back on Tuesday 29 September.
Mathieu